M6 design amendment 3g: marital projection domain law for sub-15-at-anchor persons (§2.8.2 pin correction, closes reg-5 crash)

[MaxGhenis]

M6 design amendment 3g — the marital projection domain law for sub-START_AGE-at-anchor persons

Docs-only design amendment (revision 13) adding §2.8.2g and correcting the §2.8.2 universe pin and the §2.8.9 "residual open decisions: none" claim. Closes the fifth-registration execution failure (gate_m6 candidate 1): ValueError: certified marital panel has no entry row at anchor. No code in this PR — the implementing patch is pinned in §2.8.2g and lands after ratification, as amendments 3d/3e did. No gates.yaml/runs/ contact.

Evidence base (both cited by path+field throughout §2.8.2g):

• Forensics round 2 — #42 comment 4979437110 (the dispositive decomposition)
• Grading (reg-5 execution failure) — #42 comment 4979269487

The defect

build_anchor_frame (m6_cells.py:122-140) admits every gated-start-wave person with no age filter, so minors are in the anchor. marital_panel_builder overrides start_exposure_year := anchor_wave (panel_builders.py:187) and demands the certified person_years entry row at that wave. But the certified marital person_years begin at birth_year + START_AGE (START_AGE = 15, transitions.py:111,260), so any person with anchor_wave < birth+15 has no anchor row and the builder raises (panel_builders.py:209-214). The raise prints missing[:10] (reg-5 listed 10 ids); on the realized anchor the identical mechanism trips a single uniform class of 2,850 sub-START_AGE-at-anchor children (verified read-only: 2,568/246/36 anchored 2015/2017/2019; ages 7–14; born 2001–2008).

Pin correction (uncontested)

The §2.8.2 pin claimed _valid_persons "resolves the missing-entry-row case by construction." It does not: _valid_persons (transitions.py:265-277) tests start_exposure_year ≤ censor_year on the certified start_exposure = birth+15, before the builder overrides start_exposure := anchor_wave. Person-years presence at the anchor requires anchor_wave ≥ birth_year + START_AGE. Corrected in place; §2.8.9's falsified "none" and item (iii) carry dated corrections.

The A-vs-B verdict — argued from the numbers (all verified read-only)

The 2,850 are not gated-neutral: 281 (all born 2001, age 14 at the 2015 anchor, reaching 18 by 2019) sit in the truth-side gated first_marriage.{18-29}@2019 at-risk person-years — 281 of 3,224 = 8.7 % of persons, 9.0 % of F6 weight. Zero appear in any gated event; 2,569 are gated-neutral.

	frozen v3 floor (e931c886…)	truth-side gated denominators	projected universe	report-only
A exclude-and-mark	moves → dies: symmetric exclusion drops the 281 from the truth-side first_marriage.{18-29}@2019 denominator → σ + tolerance recompute → not byte-identical. Asymmetric (projected-only) preserves the bytes but scores a 2,943-person projected vs 3,224-person truth denominator → identity-guard-rejected (support.py:311-315) + rate-biasing	moves (symmetric)	2,850 excluded	2,850
B seed-at-marital-entry (adopted)	zero — builder is projected-side only; truth side untouched	zero — 281 stay in the frozen denominator	+2,638 seeded (212 born-2008 dropped by the existing start_exposure ≤ censor filter); the 281 enter the projected denominator symmetric with truth	2,357

B is adopted; A dies on the ratified-frozen v3 floor constraint. B's pin: replace the unconditional override with start_exposure_year = max(anchor_wave, birth_year + START_AGE) — byte-unchanged for the bulk, the certified risk-set entry for the gap class.

B's named leak, adjudicated empirically (not merely "structurally near-certain")

The certified entry row at birth+15 is never_married for all 2,850 (0 exceptions), with marriage_duration/years_since_dissolution null/zero; the earliest married person-year anywhere in the class is age 19; the 11 who marry (n_marriages > 0) marry at age ≥ 19, strictly after the seed. The seed is a structural constant carrying zero holdout information → B leaks nothing. B reads the certified entry row (the same pd.isna → never_married the core applies, marital.py:100-101), not an assumed state.

Implementing patch (pinned; not in this PR)

• engine/panel_builders.py marital_panel_builder: one-line clamp at :187 (birth_year already on attrs).
• assembly.marital_step/m6_runner unchanged (identical builder signature/return).
• The discriminating test: test_m6_panel_builders.py gains a sub-START_AGE-at-anchor fixture (born 2001 anchored 2015 — the class the current born-1980–1982 fixtures lack) asserting pre-patch raise / post-patch 2016 never_married seed + holdout_ids membership, plus a born-2008 censor-drop case.
• No schema-audit/manifest/gates.yaml/runs/ delta (clamp reads only existing attrs columns).

What this does NOT change

Frozen floors byte-identical (runs/m6_holdout_floors_v{1,2,3}.json untouched; v3 sha e931c886… remains the gated contract). No gated cell definition moves. The certified engine invariant is preserved — the raise (panel_builders.py:209-214) stays; amendment 3g removes the cause (the sub-START_AGE anchor override) so the invariant holds by construction, mirroring §2.8.3a. Harness/universe-side only.

Tests

• tests/test_gate_m6_floors.py (the doc-bound test): 23 passed — all asserted design-doc strings preserved.
• tests/test_m6_panel_builders.py: 7 passed — confirms docs-only (no code changed; the current fixtures do not exercise the crash class, the §2.8.2g gap this amendment names).

Docs-only: 187 insertions, 9 deletions, one file.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
social-security-model	Ready	Preview, Comment	Jul 15, 2026 11:49am

[MaxGhenis]

Adversarial design referee — PR #209 (amendment 3g, §2.8.2g marital projection domain law)

Verdict: SPEC SOUND (ratifiable after listed fixes). 0 BLOCKING · 2 SHOULD-FIX · 3 NOTE.

I re-derived every load-bearing number independently on the real PSID panel (run-5 venv, read-only through the harness loaders — nothing written to any held worktree, runs/, or gates.yaml), and re-ran the two kills and the leak from the data rather than from the PR text. The A-kill survives on both branches and the B-leak verdict holds — the sub-START_AGE-at-anchor class reproduces exactly, symmetric-A moves the frozen tolerance 0.356 → 0.355 (byte-identity dies), the §2.8.4 identity guard fires on asymmetric-A, and all 2,850 class entry rows are never_married. The two SHOULD-FIX items are factual corrections to supporting numbers in §2.8.2g; neither revives A nor opens a leak, so they do not change the adjudicated B-over-A outcome or the pinned patch.

Independently re-derived class counts (real data)

quantity	doc	re-derived	✓
class size	2,850	2,850	✓
anchor split 2015/2017/2019	2,568 / 246 / 36	2,568 / 246 / 36	✓
birth range / age@anchor	2001–2008 / 7–14, μ10.8	2001–2008 / 7–14, μ10.8	✓
py_year_min == birth+15 (all)	yes	yes, 0 violations	✓
rows at anchor wave	0	0	✓
reg-5 crash ids = missing[:10]	(seed-0 hh side)	exact match on seed-0 household side	✓
truth first_marriage 18-29 @2019	281 / 3,224	281 / 3,224, all born 2001	✓
person / weight share of that slice	8.7% / 9.0%	8.72% / 8.97%	✓
gated events in class	0	0 (also 0 shock)	✓
born-2008 censor-drop	212	212 (only born-2008)	✓
net seeded (281 gated + 2,357 RO)	2,638	2,638	✓

The A-kill — survives (both branches)

• (i) symmetric exclusion. I reproduced the frozen v3 first_marriage.18-29|female cell byte-exact (mean 0.10667990746937188, sd 0.08307374839960044, σ 0.13521039283588324, tol 0.356), then re-ran the 100-seed household floor with the class excluded: mean/sd/σ all move and the derived tolerance shifts 0.356 → 0.355. So symmetric-A is genuinely not byte-identical → dies on the ratified-frozen v3 lock. The floor-arithmetic dependence is real and I consumed it end-to-end.
• (ii) asymmetric exclusion. The §2.8.4 identity guard (support.py:311-315) fires on real frames when the projection drops the class rows truth still carries (ValueError: symmetric presence-conditioning requires identical projection and truth person-period support); the symmetric case passes and the 281 class 2019 keys survive conditioning. Guard genuinely rejects unlike denominators.

The B-leak — verdict holds

All 2,850 certified entry rows at birth+15 are never_married (0 exceptions), marriage_duration/years_since_dissolution null/zero. Adversarial probes the PR did not run:

• Pre-15 marriages. The MH file does contain 121 pre-15 marriage-episode start ages (age hist: 14→92, 13→18, 12→6, 11→3, 10→1, 3→1), and 23 fall in the certified attrs∩anchor universe — all carried as married at their birth+15 entry. None are in the class (they are born 1928–1988), so the class's uniform never_married is real, not a coincidence, and B reads the certified row (not an assumed state) either way. No leak.
• marital.py:100-101 pd.isna semantics agree with the seed coding (never_married → state 0), consistent with the truth side.

Findings

SHOULD-FIX 1 — the "marry at age ≥19" claim is false.
§2.8.2g and the JSON amendment_3g_section state: "the 11 with n_marriages > 0 marry at age ≥ 19 (strictly after the seed)." On real data the earliest first-marriage event in the class is age 18 — person_id 2852062 (born 2004, first_marriage in 2022). Age 19 is correct only for the earliest married person-year (a consequence of _assign_state's allow_exact_matches=False, which shifts the married state to the year after the event). Also only 9 of the 11 n_marriages>0 have a dated first-marriage episode (2 are undated MH records → no changepoint → no event). The B-leaks-nothing verdict is unaffected (every event at age ≥18 is strictly after the age-15 seed; all entry states are never_married), but correct the wording to "marry at age ≥ 18 (earliest married person-year age 19); the 9 with a dated first marriage marry strictly after the seed."

SHOULD-FIX 2 — 281/3,224 = 8.7% mislabels the floor-cell perturbation.
The 3,224 (and 2,943 = 3,224−281, and 8.7% persons / 9.0% weight) is the 2019 sex-pooled at-risk slice. The actually-gated floor cell is first_marriage.18-29|female, whose denominator pools 2015–2019 and is female-only: 7,832 at-risk person-years, of which the class contributes 157 (all at 2019) = ~2.0% by rows / 2.25% by weight, not 8.7%. The A-kill is unaffected — I reproduced the frozen tolerance 0.356 and its shift to 0.355 under symmetric exclusion, so any nonzero move breaks byte-identity regardless of magnitude — but the doc overstates the gated-cell denominator move ≈4×. Recommend labeling 3,224 / 8.7% explicitly as the 2019 sex-pooled slice and citing the ~2% gated-cell (first_marriage.18-29|female, pooled) denominator move (or just "nonzero → byte-identity broken").

NOTE 1 — asymmetric-A bias direction is inverted.
"…biases the first_marriage rate downward on the projection" (line ~1066): dropping zero-event never_married at-risk person-years raises the rate. Truth-side check: excluding the class lifts the pooled female first_marriage rate 0.04847 → 0.04959 (+2.3%; events unchanged at 339, at-risk 7,832 → 7,675). Moot for the adjudication (the identity guard fires on the key mismatch first — verified), but the prose should read "upward" or drop the direction word.

NOTE 2 — missing[:10] provenance.
The reg-5 ids [6035, 7053, 10187, 23032, 23033, 39033, 39034, 40044, 40045, 41035] are missing[:10] of the seed-0 household-split side (verified identical), not of the full class — whose lowest-10 are [6035, 6038, 7050, 7051, 7052, 7053, 7054, 7055, 7214, 10036]. The doc/forensics are technically correct (the crash was in the first gate seed's household side), but "the raise reports missing[:10]" could mislead a reader into expecting the full-class lowest ids. Optional one-clause clarification.

NOTE 3 — data-quality aside (out of 3g scope). The MH file's pre-15 marriage-episode start ages include an age-3 record; 23 pre-15 persons land in-universe and are carried married at birth+15. None touch the class, so 3g is sound — worth a separate look, not a 3g blocker.

Mechanics & patch pin — all clean

• Pin correction excised. The §2.8.2 universe row no longer claims _valid_persons "resolves the missing-entry-row case by construction"; it now states the true condition (anchor_wave ≥ birth+START_AGE) and that the certified-span test runs before the anchor override. JSON builders.marital_universe carries the same correction ("CORRECTED by amendment 3g"; no residual false claim). §2.8.9 "residual open decisions: none" → "one, now closed"; item (iii) correctly distinguishes synthetic newborns (outside holdout_ids) from anchor-present minors (inside, do age into the gated band) — matches the 281 I measured.
• Patch byte-neutrality (real data). Replaying the pinned start_exposure_year = max(anchor_wave, birth+START_AGE): bulk attrs and entry person_years are byte-identical to the unpatched builder (assert_frame_equal, check_dtype=True); max() == anchor_wave for every bulk row; the class is seeded at birth+15 never_married (2,638), and the 212 born-2008 are dropped by the existing start_exposure ≤ censor filter. birth_year is on attrs (transitions.py:239-243), START_AGE from data.transitions. The certified raise (panel_builders.py:209-214) stays; callers (assembly.marital_step:279, runner) unchanged. Blast radius checked — the only other start_exposure_year consumers are marital.py:68/302 (core + fertility); the clamp moves the class's start from age 14→15, which is inert (no marital risk or fertility below 15) and gated-cell-inert.
• Discriminating-test spec accurately covers pre-patch-raise / post-patch born-2001→2016 never_married seed + holdout_ids + entry row at 2016 / born-2008 censor-drop. Confirmed the current fixtures are all born 1980–1982 (start_exposure 1995–1997) → none exercise the crash class.
• Docs-only: 1 file, +187/−9. No runs/, gates.yaml, or .py touched. Frozen floors byte-identical (v3 sha e931c886 at head). JSON block parses (single fence, revision 13). §2.8.2g header/placement mirror §2.8.3f. test_gate_m6_floors.py (23) + test_m6_panel_builders.py (7) = 30 passed at head. CI green (build, lint, pytest 3.11/3.13, Vercel); MERGEABLE / CLEAN.

Referee verification: read-only via the harness loaders on the run-5 venv; scratch confined to my own dir; my worktree removed after posting.

[MaxGhenis]

Referee fixes applied (SPEC SOUND → all 2 SHOULD-FIX + 3 NOTE addressed)

Thanks for the independent re-derivation. Every number was re-verified read-only against the run-5 venv before editing; the B-over-A adjudication, the pinned patch, and the frozen-floor byte-identity are unchanged. Fixes on m6-amendment-3g (PR stays draft for the verification round):

finding	resolution	commit
S1 — "marry at age ≥19" false	Earliest first-marriage event in the class is age 18 (person_id 2852062, born 2004, 2022); the age-19 figure is the earliest married person-year (the _assign_state allow_exact_matches=False shift). Only 9 of 11 n_marriages>0 have a dated episode. Leak conclusion unchanged — every event is strictly after the age-15 seed.	86e9354 (prose), 63f9ca3 (JSON)
S2 — 281/3,224 = 8.7% mislabels the gated cell	Relabeled as the 2019 sex-pooled slice. The gated floor cell is `first_marriage.18-29	female(female-only, pooled 2015–2019): **7,832** at-risk person-years, class contributes **157 = 2.0% rows / 2.25% weight** (all at 2019). A-kill unchanged — I state the referee-reproduced0.356 → 0.355` and that any nonzero move breaks byte-identity.
N1 — asymmetric-A bias direction	"downward" → "upward" (dropping zero-event never_married at-risk person-years raises the rate; independently confirmed the direction). Moot behind the identity guard, now stated correctly.	86e9354, 63f9ca3
N2 — missing[:10] provenance	Clarified: the 10 reg-5 ids are the lowest-id members of the crash's seed-0 household-split side, not the full class's lowest ids.	86e9354, 63f9ca3 (defect para + next-chain wording)
N3 — pre-15 MH data-quality aside	Added a one-clause out-of-scope parenthetical: the MH file carries pre-START_AGE episodes (23 land in attrs ∩ anchor, entering married at birth+15, all born 1928–1988, none in the class) — which is exactly why B reads the certified entry row rather than assuming never_married. Flagged as out of 3g scope.	86e9354

Independently re-derived, matching the referee's table: class 2,850 (2,568/246/36); event age 18 / married-py age 19 / 9-of-11 dated; female-pooled cell 7,832 with class 157 = 2.0%/2.25%; asymmetric-A direction up. JSON re-validates (revision 13, single fence); test_gate_m6_floors.py 23 + test_m6_panel_builders.py 7 = 30 passed. Docs-only; frozen floors byte-identical (v3 e931c886); no runs//gates.yaml/.py touched. Ready for the verification referee.

[MaxGhenis]

S2 propagation (follow-up to the fixes above). Completed the S2 relabel across the whole subsection so no residual pre-S2 framing remains: the A-vs-B table rows A/B, the "Symmetric, floor-inert" pin, the JSON B-description, and the JSON next-chain now all state the gated cell as first_marriage.18-29|female with the class move 157 / 2.0 % (tol 0.356 → 0.355; any nonzero move breaks byte-identity). A consistency sweep confirms the only remaining 281 / 3,224 / 8.7 % uses are the two explicitly-labeled sex-pooled 2019 slice references (which are correct). Commit 225eaa0. JSON re-validates (revision 13); test_gate_m6_floors.py 23 passed.

[MaxGhenis]

Referee verification round — commits 86e9354 / 63f9ca3 / 225eaa0 (head 225eaa0)

Verdict: VERIFIED — ratifiable. All five findings resolved, not renamed. Re-derived every number myself on the real panel (run-5 venv, read-only); re-ran the floor at this round.

finding	status	check
S1 event-vs-person-year	RESOLVED	Prose + JSON now state: earliest first-marriage event age 18 (pid 2852062, born 2004, married 2022); earliest married person-year age 19 (named as the allow_exact_matches=False shift); 9 of 11 n_marriages>0 dated; "every such event strictly after the age-15 seed → B leaks nothing." Re-derived: min event age 18, age-18 pid = [2852062], born 2004, dated 9/11, earliest married person-year 19. ✓
S2 8.7% relabeled everywhere	RESOLVED	8.7%/3,224 now labeled "sex-pooled 2019 slice" in all 6 sites (Step-2 prose, A-vs-B rows A+B, "B is adopted" para, Pins(B), JSON amendment_3g_section, JSON next). Gated cell correctly stated as first_marriage.18-29|female, 7,832 at-risk, class 157 = 2.0% rows / 2.25% weight. Re-derived exactly: 7,832 / 157 / 2.0% / 2.25%; the 2019 sex-pooled slice 3,224 / 281 / 8.72% / 8.97%. ✓
N1 bias direction	RESOLVED	"biases the rate upward … (dropping zero-event at-risk person-years raises the rate)" in prose + JSON. ✓
N2 missing[:10] framing	RESOLVED	"the 10 ids reg-5 listed are the lowest-id members of the crash's seed-0 household-split side, not the full class's lowest ids." ✓
N3 23 married-at-15 note	RESOLVED	"23 land in attrs ∩ anchor and enter married at their birth+15 — but none is in this class (born 1928–1988); out of scope for 3g." Re-derived: 23 in-universe, entry states {married: 23}, 0 in class, births 1928–1988. ✓

A-kill re-reproduced this round: the frozen v3 first_marriage.18-29\|female cell reproduces byte-exact (mean 0.10667990746937188, sd 0.08307374839960044, σ 0.13521039283588324); symmetric exclusion of the class shifts the derived tolerance 0.356 → 0.355 — the doc's stated numbers, confirmed. "Magnitude immaterial; any nonzero move breaks byte-identity" is the right framing.

Residual-framing sweep: clean. No bare 2,943, {18-29}, or @2019 remain; the only 8.7%/3,224 occurrences are the two correctly-labeled sex-pooled references (prose §2.8.2g + JSON) — the 8.7-pattern hits on lines 840/1516/1577 are §2.8.7 section refs, not S2 framing. marital_universe still carries the pin correction ("CORRECTED by amendment 3g"); residual_open_decisions = "one, closed by amendment 3g".

Mechanics at head 225eaa0: docs-only (1 file, +201/−9); frozen floors byte-identical (v3 sha e931c886); JSON block parses (single fence, revision 13); test_gate_m6_floors.py (23) + test_m6_panel_builders.py (7) = 30 passed; CI green (build, lint, pytest 3.11 + 3.13, Vercel); MERGEABLE (PR remains draft — not merging).

No new findings. The referee round is closed from my side; ratifiable.